Proceedings of the Annual Meeting of the Cognitive Science Society

Humans are remarkably adaptive instructors who adjust advice based on their estimations about a learner's prior knowledge and current goals. Inspired by prior work in rational pedagogy, we model teachers that reason about how learners will update their beliefs when given different examples, and thereby select examples that minimize expected learner error. We demonstrate that Bayesian non-parametric approaches can characterize teaching strategies in continuous domains, where traditional rational teaching models are intractable. We compare human teaching choices against our model and a variety of heuristics. Our model explains significant variance in human choices beyond a mixture of heuristics, and provides insight insight into how teachers formulate pedagogical guidance in computationally tractable ways, even in vast spaces.